Neural Machine Translation (NMT) vs Large Language Model (LLM) Translation
Neural Machine Translation (NMT) is based on a sequence-to-sequence model architecture, where an encoder encodes source language text into context vectors, and a decoder generates target language text. This architecture requires large amounts of parallel corpora (corresponding texts in source and target languages) for training, learning the mapping relationship between two languages by adjusting model weights.
The translation capability of Large Language Models (LLM) is not specifically designed, but rather a natural extension of their general language understanding abilities. When the model scale is large enough, it can automatically align multilingual semantic spaces, learning grammar rules, contextual understanding, and world knowledge. Therefore, in translation tasks, LLM translation quality often surpasses NMT, but it is also more prone to hallucinations.
The table below compares the characteristics of traditional Neural Machine Translation (NMT) engines and emerging Large Language Models (LLM) in translation scenarios.
Neural Machine Translation (NMT)
Google Translate, DeepL, Microsoft
- Extremely fast and stable
- Can handle massive concurrent requests
- Wide language support
- No context awareness, prone to losing plot continuity, especially when translating sentence by sentence
- Higher cost ($10-25 / 1M characters)
- Limited customization capabilities, can improve professionalism through glossaries
Large Language Models (LLM)
OpenAI, Claude, DeepSeek, Gemini
- Context-aware understanding, better grasp of plot and dialogue coherence
- Naturally suited for text processing tasks, with continuously improving model capabilities
- Relatively lower cost (~$0.10-5.00 / 1M characters)
- Translation speed is typically slower than NMT
- Bidirectional billing for input and output, or token-based pricing
- Limited output stability, may not perform as well as specialized NMT models in certain professional domains
Pricing Mechanism
NMT services are typically provided through REST API, using a billing model based on input character count. Some providers also offer page-based pricing schemes (for proprietary formats like doc, pdf, which are not discussed in detail here). After binding a credit card, major platforms usually provide free translation quotas ranging from 500,000 to 2 million characters per month, with charges applied for usage beyond the quota.
LLM services are also provided through REST API, but use a bidirectional billing model for input and output (Why? Because you can use simple instructions to have large language models translate large amounts of text, for example: "Please translate all 7 volumes of Harry Potter into Chinese"). Additionally, if directly connecting to large language models rather than using the platform's REST API, billing is typically based on token count rather than character count.
Given the different billing methods of these two services, in the price calculator in the next section, we approximate token-based billing to character-based billing to uniformly evaluate and compare the marginal costs of both translation methods.
Price Calculator
Based on your estimated translation volume, use the calculator below to quickly estimate monthly translation costs.
Translation Cost Calculator
Traditional NMT Services
| Provider | Price / 1M Chars | Free Tier (Monthly) | Est. Cost |
|---|---|---|---|
Google Cloud Translation | $20.00 | 500k chars | $0.00 |
Microsoft Azure Translator | $10.00 | 2M chars | $0.00 |
Amazon Translate | $15.00 | 2M chars (12 mo) | $0.00 |
DeepL API Pro | $25.00 | None | $1.49 |
Baidu General Translate | $7.00 | 1M chars (Verified) | $0.00 |
Large Language Models
| Model | Price / 1M Chars | Description | Est. Cost |
|---|---|---|---|
Claude Sonnet 4.5 | $4.500 | Literary & Nuanced | $0.2690 |
Gemini 3 Pro | $3.500 | Long Context Expert | $0.2092 |
GPT-5 | $2.813 | Advanced Agentic | $0.1681 |
GPT-5 mini | $0.563 | Fast & Affordable | $0.0336 |
Qwen-Plus | $0.400 | Balanced Performance | $0.0239 |
Character to Token Conversion (1 char ≈ tokens, ascending):
1. Western European languages (French/German/Spanish/Italian): 1 char ≈ 0.17–0.20 token
2. English: 1 char ≈ 0.25 token
3. Korean: 1 char ≈ 0.4–0.6 token
4. Japanese: 1 char ≈ 0.6–1 token
5. Chinese (Simplified/Traditional): 1 char ≈ 0.9–1.25 token
Prices are estimated based on official documentation as of late November 2025. Does not include volume discounts or subscription base fees.